Journal of Experimental Psychology: Human Perception and Performance 2005, Vol. 31, No.

3, 480 501

Copyright 2005 by the American Psychological Association 0096-1523/05/$12.00 DOI: 10.1037/0096-1523.31.3.480

Calibration, Information, and Control Strategies for Braking to Avoid a Collision

Brett R. Fajen
Rensselaer Polytechnic Institute
This study explored visual control strategies for braking to avoid collision by manipulating information about speed of self-motion. Participants watched computer-generated displays and used a brake to stop at an object in the path of motion. Global optic flow rate and edge rate were manipulated by adjusting eyeheight and ground-texture size. Stopping distance, initiation of braking, and the magnitude of brake adjustments were influenced by both optical variables, but global optic flow rate had a stronger effect. A new model is introduced according to which braking is controlled by keeping the perceived ideal deceleration, based in part on global optic flow rate, within a safe region between 0 and the maximum deceleration of the brake.

Among the most routine maneuvers performed by pedestrians, athletes, automobile drivers, and pilots is slowing down from a high speed to stop before reaching an obstacle in the path of motion. A drivers safety, and that of others, depends on the ability to control deceleration by applying the proper amount of braking pressure at just the right time. If deceleration is initiated too late or increased too gradually, eventually it will be impossible to avoid a collision. Initiating deceleration too early or increasing deceleration too rapidly may yield unnecessarily lengthy approach times or overly conservative stopping points far in front of the obstacle. On what visual information do people rely to avoid collisions with objects in the path of motion? How do people use that information to adjust deceleration from the onset of braking to the stopping point? As a person approaches an object, the optical angle subtended by that object undergoes expansion. Lee (1976) demonstrated that the time remaining until the person collides with the object (i.e., time to contact, or TTC) is approximated by the optical variable / . Since Lees discovery, has been implicated in numerous perceptionaction tasks, such as catching and hitting a ball (Bootsma & van Wieringen, 1990; Savelsbergh, Whiting, & Bootsma, 1991), detecting collisions (Andersen, Cisneros, Atchley, & Saidpour, 1999; Craig & Bootsma, 2001; Kim, Turvey, & Carello, 1993), and steering toward a target (Fajen, 2001). To control braking, Lee proposed that the first temporal derivative of ( , or tau-dot) could be maintained at a constant value in the safe range between 0 and 1.0 (hereafter, this strategy is referred to as the constant strategy). Although all values of between 0 and 1.0 will, in principle, result in a stop at the target, the deceleration profile required to maintain a particular value of will differ. For values of between 0.5 and 1.0, approach time

This research was supported by National Science Foundation Grant BCS 0236734. Thanks to Wesley Hanchar, Chris Werner, Keith Bujak, and Michael Silva for helping to collect data and to David Rom, Parthipan Pathmanapan, and Ian Grossberg for creating computer-generated displays. Correspondence concerning this article should be addressed to Brett R. Fajen, Department of Cognitive Science, Carnegie Building 308, Rensselaer Polytechnic Institute, 110 8th Street, Troy, NY 12180. E-mail: fajenb@rpi.edu 480

is short, and deceleration is low early in the approach but explodes to infinity at the end. Hence, it is impossible in practice to maintain a constant value of in this range. For values between 0 and 0.5, deceleration is finite throughout the entire approach. However, because approach duration increases dramatically for larger values of , maintaining a constant in this range is also impractical. Holding exactly at 0.5 is a special case that yields a constant deceleration approach and results in a smooth, efficient, safe stop at the object. Thus, effective use of the constant strategy would presumably involve holding close to 0.5. Kim et al. (1993) tested the sensitivity of human observers to by asking participants to make perceptual judgments about the severity of an upcoming collision. They suggested that when deceleration is adjusted to maintain at a constant value between 1.0 and 0, 0.5 functions as a critical value separating soft collisions ( 0.5) from hard collisions ( 0.5). Indeed, Kim et al. found that the mean value of separating judgments of soft and hard collisions was not significantly different from 0.5. In a response to Kim et al.s study, Kaiser and Phatak (1993) pointed out that for constant approaches, 0.5 does not have the special status attributed to it by Kim et al. Although the final deceleration required to maintain a constant approach between 1.0 and 0.5 is infinite (and, hence, unrealizable in practice), all constant approaches within the 1.0 0 range result in a stop at the target and yield the same severity of contact. Moreover, although the Kim et al. study suggests that humans are sensitive to , it does not indicate that is actually used to actively control deceleration during braking. The evidence from several earlier studies on human walking (Wann, Edgar, & Blair, 1993), automobile driving (Spurr, 1969), helicopter landing (Moen, DiCarlo, & Yenni, 1976), aerial docking in hummingbirds and pigeons (Lee, Davies, Green, & van der Weel, 1993; Lee, Reddish, & Rand, 1991), and somersaulting (Bardy & Laurent, 1998; Lee, Young, & Rewt, 1992) suggests that humans and animals do not maintain constant values of between 1.0 and 0. In retrospect, it seems unlikely that humans and animals could maintain constant values of given the finite precision and inherent limitations (e.g., perceptual thresholds, neuromuscular lags, perceptualmotor delays) of the perceptual and motor systems. This does not necessarily

CALIBRATION AND BRAKING

481

mean that is not used, only that it is not used in the manner suggested by the constant strategy.

Error-Nulling Models of Braking The Model

Yilmaz and Warren (1995; see also, Bardy & Warren, 1997) proposed that rather than holding constant, observers could adjust the brake so as to null the error ( ), which corresponds to the difference between the current value of and a margin value (denoted as m) equal to 0.5. The effectiveness of this strategy is a result of the fact that provides information about the sufficiency of the current rate of deceleration. When 0.5, holding the current rate of deceleration constant will result in a collision. Hence, deceleration is too low and the observer must increase brake pressure. When 0.5, the current rate of deceleration is too high, and the observer should decrease brake pressure to avoid stopping short of the object. Using closed-loop computergenerated displays in which observers adjusted a hand brake to stop at an obstacle in the path of motion, Yilmaz and Warren provided evidence to support this hypothesis. They found that observers tended to increase or decrease brake pressure when was less than or greater than 0.5, respectively, suggesting that 0.5 does function as a margin value for controlling deceleration. In addition, Yilmaz and Warren found only minor differences when distance cues and information about speed of selfmotion were removed, suggesting that successful braking can be achieved on the basis of alone. Despite the empirical support for the detection and use of , there are several unresolved issues regarding how the strategy could be implemented across the range of conditions that people encounter in the real world. For example, informs an observer about whether to increase or decrease deceleration but not by how much. That is, the theory of braking clearly states how the direction of brake adjustments is determined: Observers increase brake pressure when 0.5 and decrease brake pressure when 0.5. However, it is not clearly understood how the magnitude of brake adjustments is controlled. The error may be the same in two situations, but the magnitude of the change in deceleration required to null the error will depend on the observers speed and TTC with the target. For example, suppose Observer A is traveling toward a distant obstacle (24.0 m away) at a high speed of 20 m/s while decelerating at 5 m/s2, and Observer B is traveling toward a nearby obstacle (7.5 m away) at a low speed of 5 m/s while decelerating at 1 m/s2. Although is the same in both cases (0.2), the difference between the current deceleration and the deceleration that must be achieved to null is much greater for Observer A (3.33 m/s2) than it is for Observer B (0.66 m/s2). If both observers make brake adjustments proportional to , Observer A will undershoot the margin value, increasing the risk of crashing, and Observer B will overshoot the margin value, resulting in unstable and inefficient behavior. When observers are able to continuously regulate deceleration and have sufficient time to correct for undershoots and overshoots, most errors will eventually be eliminated. However, in rapid-braking conditions, overshoots and undershoots can be especially costly. To be effective, brake adjustments must be well-calibrated and finely tuned to changes in conditions. Despite the appeal of there being a single optical variable for the control of braking, many other sources of information are available

that could be used either in addition to or instead of . Andersen et al. (1999) tested the hypothesis that observers use to judge whether they will collide with or stop short of a stationary obstacle during a constant deceleration approach. Although judgments were highly correlated with , they were also influenced by factors that affect both perceived target distance (e.g., target size) and perceived speed of self-motion (e.g., edge rate). More recently, Andersen and colleagues (Andersen & Sauer, 2004; Saidpour & Andersen, in press) showed that the percentage of total trial time at which braking was controlled at different values of was affected by target size and edge rate, confirming that these effects generalize to an actively controlled braking task similar to Yilmaz and Warrens (1995) task. These effects are consistent with several other studies on timing of interceptive actions suggesting that observers rely on optical variables other than (DeLucia & Warren, 1994; Gray & Regan, 1998; Michaels, Zeinstra, & Oudejans, 2001; Rushton & Wann, 1999). Because is independent of size and speed, Andersen et al. concluded that additional sources of information are used to detect and avoid collisions with objects in the path of motion.

The Deceleration Error Model

In addition to the model, there is another error-nulling model that computes the difference between the current deceleration (dcurrent) and the ideal deceleration (dideal), which is the rate of deceleration that would bring the observer to a stop at the intended location without his or her having to make any further brake adjustments along the way. In terms of spatial variables, dideal is given by dideal v2 , 2z (1)

where v is speed and z is target distance; z/v is equal to TTC, which is optically specified by . Furthermore, speed is optically specified by the global optic flow rate (GOFR), which is the rate of optical motion of texture elements in a given visual direction, or by edge rate (ER), which is the number of texture elements that pass by a fixed reference point in the visual field (e.g., the edge of a windshield) per unit of time (Larish & Flach, 1990).1 Substituting for z/v and GOFR for v, Equation 1 can be expressed in terms of optical variables as2 dideal GOFR . 2 (2)

Perceived dideal could be compared with perceived dcurrent to make ongoing adjustments throughout the approach. That is, when dideal dcurrent, the observer should increase deceleration; when dideal dcurrent, the observer should decrease deceleration. Optical information about dcurrent is available from the rate of change of GOFR or of ER, although psychophysical studies of human sensitivity to optical acceleration (e.g., Calderone & Kaiser, 1989)
1 Speed is optically specified by GOFR as long as eyeheight is constant and by ER as long as texture density is constant. 2 One could equivalently substitute ER rather than GOFR for v. In addition, if the size of the object (r) is fixed, then dideal is also specified by a combination of and (dideal r/2 2 tan ) or a combination of GOFR and (dideal GOFR2 tan /2r).

482

FAJEN

may lead one to question the robustness with which the deceleration could be detected optically. In addition, there are several nonvisual cues known to contribute to the sense of self-motion (Harris et al., 2002) that could provide information about dcurrent. For example, the otolith organs of the vestibular system are designed to detect linear acceleration and deceleration. Cutaneous receptors on the back and bottom of a seated automobile driver are stimulated by the changing pressure applied to the skin during deceleration. In motorcycle or bicycle riding, changing air flow over the skin could also provide cues to deceleration. In most vehicles, both deceleration and brake resistance are proportional to brake displacement. Hence, feedback about dcurrent is provided by both brake resistance and brake displacement (which could be perceived proprioceptively from the position of the arm or foot on the brake). In fact, the hand brake used in Yilmaz and Warren (1995) provided resistance proportional to deceleration, raising the possibility that observers in their experiment used haptic feedback either in addition to or instead of visual feedback. Thus, dideal could be perceived visually from various combinations of optical variables, and dcurrent could be perceived from visual, vestibular, somatosensory, or haptic feedback. The difference between dideal and dcurrent (i.e., d, or the deceleration error) could then be used to adjust the brake. Because this strategy is based on the difference between dideal and dcurrent, and because this difference constitutes an error in deceleration that must be corrected, it is referred to as the deceleration error model. Unlike the error, the deceleration error informs the observer about both the direction and the magnitude of brake adjustment. Furthermore, versions of the deceleration error model that rely on perceived speed of self-motion (Equation 2) and known target size (see Footnote 1) can account for the effects of ER and target size observed by Saidpour and Andersen (in press) and Andersen and Sauer (2004).

The Constant Deceleration Model

Andersen et al. (1999) proposed a model similar to the deceleration error model, which they called the constant deceleration model, based on a comparison of two distances: the distance of the target (zs), which could be estimated from known size (r) and visual angle (zs r/2 tan ), and the estimated distance that the observer would traverse given the current speed and fixed rate of deceleration (zv v2/2dcurrent). If zs zv, then dcurrent is insufficient and must be increased to avoid colliding with the object. If zs zv, then dcurrent is excessive and should be decreased to avoid stopping too far away. With the exception of the fact that this model uses approximated rather than actual target distance, it is formally equivalent to a version of the deceleration error model (described in Footnote 2) that assumes fixed target size.3 Hence, even though the model does not explicitly compute dideal, it is treated as a version of the deceleration error model in this article. In sum, there are two existing models of braking, both of which imply that the observer acts as an error-correcting mechanism, adjusting deceleration to null a perceived error (i.e., m , or d dideal dcurrent): (a) Lees (1976) model based on the optical angle of the target and (b) the deceleration error model based on various combinations of optical variables. The existing empirical evidence is equivocal. Yilmaz and Warren (1995) concluded that observers use to control braking on the basis of comparable performance when information about target

size and speed of self-motion was removed. However, they did report a small advantage when initial TTC was short for the condition in which the ground plane was present, suggesting that information about speed may play a role under certain conditions. Andersen and colleagues found that collision judgments (Andersen et al., 1999) and braking behavior (Andersen & Sauer, 2004; Saidpour & Andersen, in press) were influenced by target size, speed of self-motion, and ER. Because is independent of target size and ER, these authors concluded that observers use additional sources of information, such as those described in their constant deceleration model or, more generally, in the deceleration error model. Among the goals of the present study were to further explore the role of speed information in visually guided braking and to resolve the apparent discrepancy between the findings of Yilmaz and Warren (1995) and those of Andersen and colleagues (Andersen et al., 1999; Andersen & Sauer, 2004; Saidpour & Andersen, in press). As in Yilmaz and Warren, observers used a hand-operated brake to actively control their deceleration, attempting to stop as close to the target as possible. To explore the role of speed information in braking, the relation between speed and GOFR and that between speed and ER were manipulated. Both GOFR and ER are proportional to ground speed, but whereas GOFR is inversely proportional to eyeheight and unaffected by texture density, ER is unaffected by eyeheight and proportional to texture density. Larish and Flach (1990; see also Ballard, Roach, & Dyre, 1998; Dyre, 1997) demonstrated that judgments of speed of self-motion are dependent on both GOFR and ER. In the present study, GOFR was manipulated by adjusting the simulated eyeheight of the observer (see Figure 1A). Because GOFR is inversely proportional to eyeheight, information about speed based on GOFR will specify either a slower or a faster speed when eyeheight is raised or lowered, respectively. ER was manipulated by scaling the size of the texture elements composing the ground plane (see Figure 1B). Because ER is proportional to ground-texture density, speed information based on ER specifies a faster or a slower speed when texture density is increased or decreased, respectively. In Experiments 1, 2, and 4, the texture density of the ground surface was rescaled with eyeheight so that GOFR and ER varied together (see Figure 1C). These variables were manipulated independently in Experiment 3. Rescaling texture density with eyeheight by adjusting the size of the tiles on the ground surface also eliminated static cues to changes in eyeheight, which would otherwise have been provided by the change in texture density and the location of the horizon on the screen. If information about speed is used to control braking, as suggested by the deceleration error model and the results of Andersen and colleagues (Andersen et al., 1999; Andersen & Sauer, 2004; Saidpour & Andersen, in press), then observers should brake harder than necessary when eyeheight is lowered and more weakly
According to the deceleration error model, observers should increase deceleration when perceived dideal is greater than perceived dcurrent. If dideal is estimated on the basis of v2/2zs, where v is perceived speed and zs is perceived distance based on known size, then observers should increase deceleration when v2/2zs dcurrent. Multiplying both sides by zs/dcurrent yields v2/2dcurrent zs. Substituting zv for v2/2dcurrent (from Andersen et al., 1999) yields zv zs, which is the same as the conditions in which observers should increase deceleration according to Andersen et al.s (1999) constant deceleration model.
3

CALIBRATION AND BRAKING

483

Figure 1. A: Global optic flow rate (GOFR) is the angular velocity of texture elements in a given visual direction ; it is inversely proportional to eyeheight (EH) and invariant across changes in texture density. GOFR was manipulated by varying EH. B: Edge rate (ER) is the number of texture elements per unit of time that pass by a reference point in a given visual direction ; it is invariant across changes in EH and proportional to texture density. ER was manipulated by varying texture density. C: Combined effects of changing EH and texture density on GOFR and ER. When the scale of the ground plane is adjusted with EH, as it was in Experiments 1, 2, and 4, GOFR and ER vary together.

than necessary when eyeheight is raised. However, if observers do not use information about speed to adjust braking, as suggested by Lees (1976) hypothesis and the results of Yilmaz and Warren (1995), then performance should be unaffected by the manipulation of eyeheight.

The Significance of Maximum Deceleration

Although there is disagreement about whether observers rely on or on cues to distance, speed, and deceleration, it is generally assumed that braking is controlled by making adjustments to null

the deceleration error (i.e., the difference between dcurrent and dideal). However, there is an important aspect of visually guided braking that cannot be captured by the or deceleration error models or by any other error-nulling model. The problem with error-nulling models is illustrated below using the model as an example, but the deceleration error model encounters the same problem. Imagine a driver moving on the highway at a constant speed toward a toll booth. According to the model, the driver should increase brake pressure because , which is always equal to 1.0 when approaching an object at a constant speed, specifies that

484

FAJEN

dcurrent is insufficient and that it is necessary to increase brake pressure until 0.5. Canceling the error will null the difference between dcurrent and dideal, guiding the driver at a constant deceleration to a safe stop at the toll booth. But of course, drivers do not need to null the deceleration error until the very last moment. A cautious driver might hit the brakes hard immediately upon recognition of the toll booth and gradually creep up to the stopping point (see Figure 2A). A more aggressive driver might wait before initiating deceleration and then slam on the brakes at the last possible moment (see Figure 2B). Within the error-nulling framework, one would simply say that the two drivers nulled the error in different ways. The model makes no specific predictions about when or how to cancel the error as long as a driver does so before reaching the obstacle, so neither driver did anything that contradicts the model. Nevertheless, it is apparent that canceling the error throughout the entire approach is not a necessary condition for successful performance. Canceling the error (i.e., nulling the difference between dcurrent and dideal) is a sufficient condition but not a condition that is always possible to satisfy. If the aggressive driver in the example above waits too long before starting to brake or increases brake pressure too gradually, at some point dideal will exceed the maximum deceleration (Dmax), and the driver will be unable to cancel the error or avoid a collision (see Figure 2C). So Dmax introduces a critical constraint on successful performance. As long as dideal Dmax, it is still possible to avoid a collision by increasing brake pressure. But if dideal Dmax, it is no longer possible to stop safely. Maximum deceleration defines an action boundary in that it separates situations in which it is still possible to stop from situations in which it is no longer possible to stop. Observers do not need to be sensitive to the sufficiency of their dcurrent, but they must be sensitive to the location of the action boundary. Unfortunately, error-nulling models (including the and deceleration error models) cannot capture this aspect of braking because the optical variables on which observers presumably rely to control braking specify the sufficiency of the observers dcurrent. These variables provide no information about how close the observer is to exceeding Dmax. So an observer relying on would be completely insensitive to the location of the action boundary. Another goal of this study was to demonstrate that observers are sensitive to the location of the action boundary. If they are, then they should be more likely to increase brake pressure when dideal is close to the maximum deceleration of the brake. Of course, larger values of dideal are likely to be accompanied by larger differences between dideal and dcurrent. However, it is possible to isolate the influence of dideal by focusing only on situations in which the difference between dideal and dcurrent is close to 0. Because the error is close to 0.5 and the deceleration error is close to 0 in these situations, any effect of dideal can be attributed to an observers sensitivity to Dmax. Sensitivity to the action boundary cannot be captured by any error-nulling model, and so such findings would lead to rejection of the error-nulling assumption and necessitate an entirely different kind of model of visually guided braking.

Figure 2. Current and ideal deceleration (upper panels) and (lower panels) for sample trials illustrating a conservative style (A), an aggressive style (B), and a crash (C). The boundary between the white and gray regions at 7.0 m/s2 is determined by the maximum deceleration of the brake.

Experiment 1
Experiment 1 was an extension of Yilmaz and Warren (1995) in three new directions. First, a shorter range of initial TTC values (2.5 4.5 s) was used to further explore the ground-plane advantage for the shortest initial TTC condition (3 s) used by Yilmaz and Warren. Second, a third condition was added to the ground and air conditions used by Yilmaz and Warren one in which the ground

CALIBRATION AND BRAKING

485

plane was present but eyeheight varied randomly on each trial between one half and twice the standard eyeheight. To distinguish between the two ground-plane conditions, the ground condition used by Yilmaz and Warren is referred to as the groundfixed-EH condition, and the new condition in which eyeheight varied randomly is referred to as the groundvariable-EH condition. Thus, it is possible to compare performance when information about speed provided by GOFR and ER is present (groundfixed-EH condition) with performance when such speed information is absent (air condition) or unreliable (groundvariable-EH condition). Third, an analysis was conducted of the likelihood of increasing brake pressure as a function of dideal to determine whether observers were sensitive to the action boundary defined by the maximum deceleration of the brake.

Method
Participants. Twelve people (2 women and 10 men) volunteered to participate in the experiment. Their average age was 21.2 years, and all had normal or corrected-to-normal vision. All but one participant had a valid drivers license, and their mean number of years of driving experience was 4.8 (SD 3.7 years). Displays and apparatus. Displays were created using OpenGL and generated by a Dell Precision 530 workstation equipped with a 1.7-GHz Intel Xeon processor and an NVIDIA Quadro2 graphics card. Displays were rear projected onto a screen at a resolution of 1,024 768 pixels and a frame rate of 60 Hz, and they were viewed from a distance of approximately 1 m. The borders of the projection screen and the surrounding walls were covered with a black felt fabric to reduce the salience of the screen frame.

Braking was controlled using a Microsoft Sidewinder Force Feedback 2 joystick. Observers increased deceleration by pulling the joystick from the neutral center position. The brake was programmed so that deceleration was proportional to joystick position. Deceleration ranged from 0 m/s2 in the neutral center position to 7 m/s2 ( 0.7 g) in the full down position. The maximum displacement of the joystick was approximately 10 cm. Joystick position was sampled at a rate of 60 Hz and used to update the display with a loop time of 1 frame. The joystick was positioned in a comfortable location on a small table in front of the observer and off to the side of his or her preferred hand. The displays simulated observer movement along a linear path toward three red-and-white, octagonal stop signs (see Figure 3). The sky was light blue, and a gray cement-textured ground surface was present in the groundfixed-EH and groundvariable-EH conditions but not in the air condition. In the groundfixed-EH condition, the observers viewpoint was 1.10 m (or 1 standard eyeheight unit [e]) above the ground surface. In the groundvariable-EH condition, the observers viewpoint varied between 0.55 m and 2.20 m (0.52.0 e). The tiles composing the ground surface were rescaled to eyeheight on each trial in the groundvariable-EH condition so that the ground surface would appear the same. One stop sign was positioned on the observers simulated path of motion, and the other two were positioned on the right and the left, respectively. The radius of the stop sign varied randomly on each trial between 0.25 and 0.50 m, and the distance between stop signs was always 4 times the radius. The center of each stop sign was at the same height as the observers viewpoint. To eliminate cues to changes in eyeheight, the signs were not anchored to the ground by a post but, rather, appeared to be floating above the surface. There were five initial distances (40.0, 42.5, 45.0, 47.5, and 50.0 m) and five initial TTCs (2.5, 3.0, 3.5, 4.0, and 4.5 s). Initial speed was determined

Figure 3.

Sample frame from the display used in the groundfixed-EH (eyeheight) condition of Experiment 1.

486

FAJEN

by dividing the initial distance by the initial TTC on each trial, resulting in 24 different initial speeds ranging from 8.9 to 20.0 m/s. Procedure. Observers sat in an adjustable chair in front of a large (1.81.2-m) rear-projection screen. The screen was located approximately 1 m in front of the observers viewpoint so that it encompassed 85 of the field of view in the horizontal dimension and 62 in the vertical dimension. Prior to the experiment, the seat was adjusted so that the observers seated eyeheight was approximately 1.10 m, which was equal to the simulated eyeheight used in the groundfixed-EH condition. Observers were instructed to use the joystick as a hand-operated brake to stop as close as possible to the stop signs. As in Yilmaz and Warren (1995), observers were given specific instructions to encourage smooth braking. They were told to imagine that they were driving an actual automobile, in which sudden or jerky brake adjustments would be uncomfortable and unsafe. They were also discouraged from waiting until the last possible moment before initiating full brake pressure. Trials were blocked by environment, and the three blocks were presented in a counterbalanced order. Prior to the beginning of the experimental session, observers completed an additional block of 100 trials in the groundfixed-EH condition. This condition was chosen for the practice block because it served as the standard condition against which performance in the other conditions could be compared. Trials were initiated by moving the joystick to the neutral, zerodeceleration position and pressing the trigger button on the joystick. The scene appeared, and simulated motion toward the stop signs began immediately. Displays ended when observers came to a stop. Even if an observer collided with the stop sign, the trial continued until speed was equal to 0. To indicate that a collision occurred, an audio signal (a beep) lasting 1 s was played on impact with the stop sign. The final frame was displayed for 1 s before the intertrial screen appeared. Design. The design for Experiment 1 was 3 (environment) 5 (initial distance) 5 (initial TTC), with all variables within participant. Trials were blocked by environment, and block order was counterbalanced across participants. Trials within each block were presented in a random order, and there were four repetitions per condition for a total of 100 trials per block and 300 trials per session (not including the block of practice trials). The entire experiment lasted approximately 1 hr. Data analysis. During each trial, the computer recorded the observers simulated distance from the target (z), speed (v), and deceleration (d) at a sampling rate of 60 Hz. These values were subsequently used to compute and for each frame on the basis of the equations in Lee (1976; see also, Yilmaz & Warren, 1995). Final stopping distance was determined by the distance from the observer to the target in the final frame (i.e., when speed equaled 0), regardless of whether the observer stopped before or crashed into the target. To determine whether observers were more likely to increase deceleration as dideal approached Dmax, the deceleration time series from each trial was sampled at intervals of 200 ms (12 frames), starting at the beginning of the trial. At each sampled frame, dideal, the deceleration error ( d), and the direction of the brake adjustment were calculated. To isolate the influence of dideal, only samples at which d was between 1.0 and 1.0 m/s2 were used in this analysis. The direction of the brake adjustment at each sampled frame was determined by calculating the change in deceleration over a 200-ms window centered on the frame. If deceleration increased or decreased by more than 0.03 m/s2 between the beginning and end of the 200-ms window, the direction of brake adjustment was recorded as an increase or a decrease, respectively. In addition, the direction of adjustment was recorded as an increase whenever deceleration was held at 7.0 m/s2 and as a decrease whenever deceleration was held at 0 m/s2 from the beginning to the end of the window. Otherwise, the direction of brake adjustment was recorded as a hold. The data were binned by dideal into 10 equally spaced bins between 1.0 and 6.0 m/s2, and the percentage of increases in each bin was calculated.

Results and Discussion

For the analyses of constant and variable error in final stopping distance and percentage of crashes, the data were collapsed across initial distance and entered into separate 3 (environment) 5 (initial TTC) analyses of variance (ANOVAs), one for each dependent variable. Constant error in final stopping distance. The ANOVA indicated significant main effects of environment, F(2, 22) 14.36, p .01, and initial TTC, F(4, 44) 70.77, p .01. The Environment Initial TTC interaction was significant as well, F(8, 88) 19.26, p .01 (see Figure 4A). The simple main effect of environment was significant at the 2.5- and 3.0-s initial TTC levels, and planned comparisons yielded significant differences between the groundfixed-EH and air conditions at the two shortest initial TTC levels (2.5 and 3.0 s). The differences between the groundfixed-EH and groundvariable-EH conditions did not reach significance. These results indicate that the source of the interaction was the difference between the groundfixed-EH and air conditions at small initial TTC values. Variable error in final stopping distance. The main effects of environment, F(2, 22) 13.80, p .01, and initial TTC, F(4, 44) 15.10, p .01, and the Environment Initial TTC interaction, F(8, 88) 3.81, p .01 (see Figure 4B), were all significant. The simple main effect of environment was significant at the 2.5-, 3.0-, and 3.5-s initial TTC levels, and planned comparisons revealed differences between the groundfixed-EH and air conditions at the 2.5-, 3.0-, 3.5-, and 4.0-s initial TTC levels and differences between the groundfixed-EH and ground variable-EH conditions at the 2.5-s initial TTC level. Thus, unlike in the analysis of constant error, the groundvariable-EH condition also contributed to the interaction. Percentage of crashes. The main effects of environment, F(2, 22) 10.21, p .01, and initial TTC, F(4, 44) 86.60, p .01, and the Environment Initial TTC interaction, F(8, 88) 6.66, p .01 (see Figure 4C), were all significant. The simple main effect of environment was significant at the 2.5-s initial TTC level only. Planned comparisons revealed significant differences between the groundfixed-EH and air conditions at the 2.5-s initial TTC level only. None of the differences between the ground fixed-EH and groundvariable-EH conditions were significant.4 Eyeheight and final stopping distance. Additional analyses were conducted to better understand the effects of eyeheight on
The analysis of percentage of crashes is included here for the purpose of comparison with Yilmaz and Warren (1995). However, it is important to note that colliding with the stop sign does not indicate failure to perform the task, because the instructions in both the present study and Yilmaz and Warrens were to stop as close as possible to the stop sign. Observers were not explicitly told to stop close without colliding because it was necessary to provide a fair test of the and deceleration error models, both of which describe how an observer stops at an obstacle rather than just before it. Although the braking task performed by automobile drivers requires that they stop before reaching an obstacle, other braking tasks (e.g., docking) require drivers to make soft, controlled contact. Thus, the task used in this study is not unrepresentative of normal braking tasks. In Experiments 2, 3, and 4, analyses of percentage of crashes are not reported. As a methodological note, Andersen and colleagues (Andersen & Sauer, 2004; Saidpour & Andersen, in press) have found that the overall percentage of crashes can be reduced by simulating self-motion for a brief period (6 s) before observers are allowed to initiate braking.
4

CALIBRATION AND BRAKING

487

final stopping distance. If observers are influenced by information about speed provided by GOFR and ER, then they should be more likely to overshoot dideal when eyeheight is low, and they may stop short of the target. When eyeheight is high, observers should undershoot dideal, and they may stop closer to or behind the target. To explore this relation, a linear regression analysis was performed on final stopping distance in the groundvariable-EH condition with eyeheight as an independent variable. The slope of the regression line provided an indication of the direction of the influence of eyeheight, and r2 provided a measure of the strength of this influence. Separate analyses were run for each observer. Overall, the mean slope was 2.73 (SD 1.23), and the mean r2 was 0.11 (SD 0.06). The relation between final stopping distance and eyeheight was significant ( p .05) for 11 out of 12 (91.7%) of the observers. Further analyses revealed that both slope and r2 were affected by initial TTC (see Figure 5). At the shortest initial TTCs, the slope was quite steep, and the relation between final stopping distance and eyeheight was strong. As initial TTC increased, the slope became more shallow and the relation weaker. One-way ANOVAs indicated significant effects of initial TTC on both

Figure 4. Constant error in stopping distance (A), variable error in stopping distance (B), and percentage of crashes (C) for each initial time to contact in the groundfixed-EH (eyeheight), air, and groundvariable-EH conditions of Experiment 1. Figure 5. Mean slope and r2 from the regression analysis of eyeheight (EH) and final stopping distance in the groundvariable-EH condition of Experiment 1. Error bars represent plus or minus 1 standard error.

488

FAJEN

slope, F(4, 44) 16.39, p .01, and r2, F(4, 44) 2.67, p .05. Thus, final stopping distance was influenced by eyeheight in the variable-EH condition, but the influence of eyeheight depended on initial TTC. It is likely that the influence of eyeheight was stronger when initial TTC was shorter because observers had less time to correct for overshoots and undershoots. When initial TTC was longer, there was more time to correct for inaccurate braking adjustments, which could account for the weaker relation between eyeheight and final stopping distance. The analysis of all three success measures consistently indicated that information about speed provided by GOFR and ER influences braking, particularly at small initial TTC values. These findings replicate those of Yilmaz and Warren (1995) and extend their study in two new directions. First, the ground-plane advantage was further magnified at the shortest initial TTC values, suggesting that this effect deserves more attention. Second, the results indicated that performance was also affected when information about speed provided by GOFR and ER was present but unreliable. When eyeheight was lower than normal, observers tended to stop short of the target. When eyeheight was above normal, observers often collided with the target. These findings are difficult to resolve with Lees (1976) original hypothesis, in which GOFR and ER do not play a role. The effects of eyeheight are consistent with versions of the deceleration error model that rely on speed information, but neither model can account for observers sensitivity to the action boundary defined by Dmax. Ideal deceleration analysis. Figure 6 shows the percentages of brake adjustments that resulted in an increase in deceleration as a function of dideal for all three conditions. Recall that to isolate the influence of dideal, only those frames for which the deceleration error was between 1.0 and 1.0 m/s2 were used. In other words, was approximately equal to 0.5, and d was approximately equal to 0 in all 10 dideal bins. Figure 6 clearly indicates that the likelihood of increasing brake pressure was affected by dideal. When dideal was close to 0, observers were just as likely to increase deceleration as they were to maintain or decrease deceleration.

When dideal was close to Dmax, however, observers almost always increased deceleration. The mean slope of the line that best fit the data for each observer was 7.8 (SD 3.8) in the groundfixed-EH condition, 7.7 (SD 3.1) in the air condition, and 7.5 (SD 2.7) in the groundvariable-EH condition.5 As confirmed by t tests, the slopes were significantly greater than 0 in the groundfixed-EH condition, t(10) 6.83, p .01; the air condition, t(10) 8.27, p .01; and groundvariable-EH condition, t(10) 9.09, p .01. Because the deceleration error was between 1.0 and 1.0 m/s2 for all frames used for this analysis, both the and deceleration error models predict that regardless of dideal, observers should maintain brake pressure or be just as likely to decrease brake pressure as they are to increase it. So these findings cannot be explained by the or deceleration error models. One might initially be surprised to see that the curves for all three conditions overlapped so closely. On the one hand, because dideal is specified in part by GOFR and ER, one might expect observers to be less sensitive to the location of the action boundary when speed information provided by GOFR and ER was either absent (in the air condition) or manipulated (in the ground variable-EH condition). On the other hand, when eyeheight was varied randomly in the groundvariable-EH condition and when the ground surface was removed in the air condition, one might expect observers to overestimate dideal as often as they underestimate it. Even if there was an effect of GOFR and ER, it would have cancelled out when the data were averaged within each condition. So, in fact, it is not particularly surprising that the curves overlapped. However, the mean r2 was slightly higher (although not significantly so) in the groundfixed-EH condition (0.62) than it was in the air (0.52) and groundvariable-EH (0.52) conditions. In sum, the results of Experiment 1 demonstrate that braking is influenced when information about speed provided by GOFR and/or ER is absent or unreliable, especially when there is less time to correct for brake-adjustment errors. Observers tended to stop short of the target when GOFR and ER were higher than normal, and they were more likely to crash into the target when GOFR and ER were lower than normal. The influence of GOFR and ER is not predicted by Lees (1976) original strategy, which relies entirely on variables defined by the optical angle of the obstacle and its derivatives. Versions of the deceleration error model that rely on GOFR or ER to estimate dideal (see Equation 2) do predict an influence of speed information. However, both models assume that observers make adjustments to null the error between dcurrent and dideal on the basis of the optical variables that specify the sufficiency of dcurrent. In Experiment 1, observers reliably increased deceleration when dideal was close to Dmax, even when the deceleration error was close to 0 (i.e., when was close to 0.5). This shows that observers are sensitive to the action boundary defined by Dmax. Optical variables that specify the sufficiency of dcurrent, such as those in the and deceleration error models, provide no information about dideal relative to Dmax. The only way to account for observers sensitivity to the action boundary is to reject the assumption, on which both models are based, that observers make brake adjustments to null the deceleration error.
Data from 1 observer were eliminated from this analysis because there were too few data points to calculate a reliable measure of the percentage of increase in one of the dideal bins.
5

Figure 6. Percentages of increase in brake pressure as a function of ideal deceleration in the groundfixed-EH (eyeheight; open circles), air (open squares), and groundvariable-EH (open triangles) conditions of Experiment 1.

CALIBRATION AND BRAKING

489 Experiment 2

A New Model of Braking

If keeping dideal within the safe region is the necessary and sufficient condition for successful performance, then perhaps observers do not adjust brake pressure to null the difference between dcurrent and dideal but, rather, to keep dideal within the safe region between 0 and Dmax. An observer using a brake with a maximum deceleration will learn that values of dideal that are less than Dmax can be kept within that region by adjusting the brake; that is, as long as dideal is below the boundary, the observer can always apply more brake pressure to keep dideal from exceeding the boundary. This is illustrated in Figures 2A and 2B, both of which show successful stops despite the fact that the observer never actually nulled the deceleration error. The ideal deceleration increased toward the boundary at the beginning of both trials, but the observers made adjustments to keep it from exceeding Dmax. According to the new model, each brake adjustment in these sample trials was made to keep dideal from escaping the safe region defined by the action boundary. An equivalent way of describing this strategy is to adjust brake pressure so that it is always still possible to stop within the limits of ones braking capabilitiesthat is, to adjust brake pressure so that safe stopping is always afforded. To distinguish the proposed model from the and deceleration error models, it is hereafter referred to as the affordance model of braking, borrowing the term affordance, originally coined by Gibson (1979), to provide a description of the environment in terms of its possibilities for action. The significance of affordance theory is explored in more detail in the General Discussion. One advantage of the affordance model is that it accounts for why observers reliably increase brake pressure when dideal approaches Dmax, as shown in Figure 6. Even when the deceleration error is close to 0, it is extremely risky to keep dideal near Dmax. If dideal is slightly underestimated, for example, then it may exceed Dmax, in which case collision is unavoidable. The model also accounts for the role of GOFR and ER reported in Experiment 1. Recall from Equation 2 that dideal is optically specified by a combination of GOFR or ER and (i.e., dideal GOFR/2 , or dideal ER/2 ). If observers rely on GOFR/2 or ER/2 to perceive dideal, then performance should have been best in the groundfixed-EH condition because GOFR and ER were present and reliable. When GOFR and ER were absent, as was the case in the air condition, performance should have been poorer because observers would have been forced to use another, less reliable variable that did not involve GOFR and ER. Observers should also have performed poorly in the groundvariable-EH condition because manipulations of GOFR and ER would have resulted in overestimation and underestimation of dideal. The relation between eyeheight and final stopping distance in the groundvariable-EH condition was consistent with the pattern of results that one would expect if observers overestimated and underestimated dideal. Lastly, when observers make adjustments to keep dideal within the safe region on both sides, as they often do, the appearance of error nulling actually emerges automatically. In sum, the proposed affordance model of braking can account for the fact that observers reliably increase brake pressure when dideal is close to Dmax, for the influence of GOFR and ER, and for the appearance of errornulling brake adjustments reported in previous studies (e.g., Yilmaz & Warren, 1995).

Experiment 1 demonstrated that braking is influenced by GOFR and/or ER, contrary to Lees (1976) original model. However, the design of Experiment 1 did not permit the kind of analyses that are necessary to understand how ongoing brake adjustments are influenced by GOFR and ER. To illustrate how these optical variables might influence ongoing brake adjustments, consider the consequences of increasing GOFR by lowering eyeheight. Observers relying on GOFR/2 to perceive dideal will overestimate dideal and should initiate deceleration earlier and brake harder to keep perceived dideal within the safe region. The opposite pattern of results would be expected when eyeheight is raised. In Experiment 2, eyeheight was manipulated again, but this time as a randomly presented, within-participant independent variable with three levels (0.55, 1.10, and 2.20 m) to allow for the possibility of measuring the influence of GOFR and ER on the timing and magnitude of individual brake adjustments. If observers make brake adjustments to keep perceived dideal (based on GOFR/2 or ER/2 ) within the safe region, they should initiate deceleration at a lower value of dideal when eyeheight is 0.55 m and at a higher value of dideal when eyeheight is 2.20 m. Eyeheight should also influence the magnitude of individual brake adjustments throughout the entire approach. The magnitude of individual brake adjustments was scaled to the magnitude of the brake adjustment required to null the deceleration error. For example, if dideal is 2.5 m/s2, and the observer increases deceleration from 1.5 to 3.0 m/s2, then the scaled magnitude of adjustment would be 1.5/1.0 1.5. That is, the observer overshot dideal by 50%. When eyeheight is 0.55 m, observers should be more likely to overshoot dideal when they are increasing brake pressure to keep perceived dideal within the safe region. When decreasing brake pressure, they should be more likely to undershoot dideal. Likewise, when eyeheight is 2.20 m, observers should be more likely to undershoot dideal when they are increasing brake pressure and to overshoot dideal when they are decreasing brake pressure.

Method
Participants. Twelve undergraduate students (7 women and 5 men) volunteered to participate in the experiment. Their average age was 20.0 years, and all had normal or corrected-to-normal vision. All participants had a valid drivers license, and their mean number of years of driving experience was 3.9 (SD 1.6 years). Displays and apparatus. As in Experiment 1, there were five initial distances (40.0, 42.5, 45.0, 47.5, and 50.0 m) and five initial TTCs (2.5, 3.0, 3.5, 4.0, and 4.5 s). In addition, eyeheight (0.55, 1.10, or 2.20 m) was manipulated as a within-subject independent variable, and the textured ground surface was present on all trials. Procedure. The procedures were also similar to those used in Experiment 1. In this experiment, however, there was only one block of 300 trials. Observers were asked to take a short break after the 100th and the 200th trial. Prior to the beginning of the experimental session, observers completed an additional block of 100 practice trials with eyeheight fixed at 1.10 m. Design. The design for Experiment 1 was 3 (eyeheight) 5 (initial distance) 5 (initial TTC), with all variables within subject. Trials were presented in a completely random order, and there were four repetitions per condition for a total of 300 trials per session (not including the block of practice trials). The entire experiment lasted approximately 1 hr. Data analysis. In addition to the analyses used in Experiment 1, detailed analyses of the individual brake adjustments made throughout

490

FAJEN

each trial were also performed. Brake adjustments were identified using a Matlab program written to find transitions points (i.e., starts, stops, and reversals) in the deceleration time series. Starts were recorded when the observer increased or decreased deceleration by at least 0.03 m/s2 for 200 ms (12 frames) or more. Brake adjustments that resulted in changes in deceleration of less than 0.03 m/s2 or that lasted less than 200 ms were not counted. Stops were recorded when the observer stopped moving the brake such that deceleration was held constant ( 0.03 m/s2) for at least 200 ms. Reversals were recorded when the observer reversed the direction of motion of the brake such that deceleration at the reversal point was at least 0.03 m/s2 greater than (for a -shaped reversal) or less than (for a U-shaped reversal) deceleration at the beginning and end of a 200-ms window surrounding the reversal point. Individual brake adjustments were identified by looking for starts followed by stops or reversals and for reversals followed by stops or reversals. The scaled magnitude of each brake adjustment was determined by dividing the difference between deceleration at the beginning and end of the adjustment by the difference between dcurrent and dideal at the beginning of the adjustment.

Results and Discussion

As in Experiment 1, data were collapsed across initial distance and entered into separate 3 (eyeheight) 5 (initial TTC) ANOVAs for the analysis of constant and variable error in final stopping distance. Constant error in final stopping distance. The ANOVA indicated significant main effects of eyeheight, F(2, 22) 38.30, p .01, and initial TTC, F(4, 44) 45.23, p .01, and a significant Eyeheight Initial TTC interaction, F(8, 88) 22.07, p .01 (see Figure 7A). The simple main effect of eyeheight was significant at all levels of initial TTC, and planned comparisons revealed significant differences between all three eyeheight conditions, with the exception of the 1.10- and 2.20-m conditions at the 4.5-s initial TTC level. Variable error in final stopping distance. The main effect of eyeheight, F(2, 22) 3.54, p .06, did not reach significance. The main effect of initial TTC, F(4, 44) 21.35, p .01, and the Eyeheight Initial TTC interaction, F(8, 88) 7.05, p .01 (see Figure 7B), were both significant. The simple main effect of eyeheight was significant at the 2.5-, 3.5-, and 4.0-s initial TTC levels. Note that variable error did not increase as initial TTC decreased in the 0.55-m eyeheight condition. It seems likely that observers tended to apply full brake pressure in this condition because initial speed was high and eyeheight was low, resulting in an extremely high GOFR. Because Dmax was fixed, this would have led to lower variability in final stopping distance compared with trials on which brake pressure was adjusted throughout. The analyses of constant and variable error in final stopping distance confirm that braking was influenced by eyeheight in the design used in Experiment 2. The main goal of this experiment was to determine the effects of eyeheight on initiation of braking and on the magnitude of brake adjustments. Initiation of braking. Mean dideal at the onset of braking is plotted as a function of initial TTC for the 0.55-, 1.10-, and 2.20-m eyeheight conditions in Figure 8. The main effects of initial TTC, F(4, 22) 1,379.02, p .01, and eyeheight, F(2, 22) 12.02, p .01, were significant, but the Initial TTC Eyeheight interaction was not (F 1). The simple main effect of eyeheight was significant at all five levels of initial TTC. The results are consistent with the prediction that the onset of braking should be affected by manipulations of GOFR and ER, which would result in overestimation and underestimation of dideal.

Figure 7. Constant (A) error and variable (B) error in stopping distance for each initial time to contact in the 0.55-, 1.10-, and 2.20-m eyeheight conditions of Experiment 2.

Magnitude of brake adjustments. Separate analyses of the magnitude of initial and subsequent brake adjustments were conducted. For initial brake adjustments, the main effects of eyeheight, F(2, 22) 9.30, p .01, and initial TTC, F(4, 44) 3.90, p .05, were significant, but the Eyeheight Initial TTC interaction was not, F(8, 88) 1.09, p .37. Figure 9A indicates that observers tended to overshoot the required initial brake adjustment by 20% 40% when eyeheight was low, with the percentage of overshoot unaffected by initial TTC except in the 4.5-s condition. When eyeheight was high, initial brake adjustments slightly overshot or matched the required initial brake adjustment at short initial TTC values and undershot the required initial brake adjustment at longer initial TTCs. For subsequent brake adjustments, separate analyses were performed for those that resulted in an increase in deceleration and those that resulted in a decrease in deceleration. The data were also binned according to TTC at the onset of brake adjustment in four intervals of 1.0 s. That is, the first interval included all brake

CALIBRATION AND BRAKING

491

adjustments initiated when TTC was equal to or less than 1.0 s, the second interval included brake adjustments initiated when TTC was greater than 1.0 s and less than 2.0 s, and the third interval included brake adjustments initiated when TTC was greater than 2.0 s and less than 3.0 s. The fourth interval included all brake adjustments initiated when TTC was greater than 3.0 s. For brake adjustments that resulted in an increase in deceleration, the main effects of eyeheight, F(2, 18) 11.44, p .01, and TTC, F(3, 27) 9.47, p .01, were significant.6 The Eyeheight TTC interaction, F(6, 54) 1.49, p .24, was not significant. The pattern of results for increases was similar to the pattern for initial brake adjustments (see Figure 9B). Observers tended to overshoot the required brake adjustment when eyeheight was low, and they tended to match or undershoot the required brake adjustment when eyeheight was high. Also, observers made larger brake adjustments (relative to the required adjustment) when TTC was short, which probably reflects a safety bias to avoid collision with the stop sign. Although the Eyeheight TTC interaction was not significant, it is interesting to note that observers tended to match or overshoot the required change in deceleration when TTC was short, even in the high-eyeheight condition. This pattern suggests that the influence of GOFR and ER was weaker when TTC was short. However, speed information was not disregarded entirely the magnitude of overshoot was still greater in the low-eyeheight condition. Because there were relatively few brake adjustments that resulted in a decrease, it was necessary to collapse the data across TTC. The mean scaled magnitude of brake adjustments that resulted in a decrease was well below 1.00 in all three conditions (M 0.61, 0.70, and 0.63 for the low-, normal-, and higheyeheight conditions, respectively), and a one-way ANOVA on eyeheight was not significant (F 1). Thus, brake adjustments that resulted in a decrease in deceleration were not influenced by GOFR or ER, and they tended to undershoot the required decrease in deceleration. These results probably reflect a bias not to release the brake too much so as to avoid colliding with the stop sign.
Figure 9. A: Scaled magnitude of initial brake adjustment for each initial time to contact in the 0.55-, 1.10-, and 2.20-m eyeheight conditions of Experiment 2. B: Scaled magnitude of subsequent brake adjustments as a function of time to contact at the onset of adjustment in the 0.55-, 1.10-, and 2.20-m eyeheight conditions of Experiment 2.

Taken together, the analyses of the initiation of braking and the magnitude of initial and subsequent brake adjustments explain the effect of GOFR and ER on final stopping distance. Observers stopped further in front of the stop signs when eyeheight was lower than normal because braking was initiated earlier and there was a tendency to brake harder. When eyeheight was higher than normal, observers stopped closer to or behind the stop signs because braking was initiated later and brake adjustments were weaker. These results are consistent with the predictions of the affordance model that observers make adjustments to keep the perceived dideal (based on GOFR/2 or ER/2 ) within the safe region.

Figure 8. Mean ideal deceleration at the onset of braking for each initial time to contact in the 0.55-, 1.10-, and 2.20-m eyeheight conditions of Experiment 2.

Data from 2 observers were excluded because there were too few adjustments in at least one bin to compute a reliable estimate of brakeadjustment magnitude.

492 Experiment 3

FAJEN

Under normal conditions, adjusting eyeheight changes GOFR but not ER because the same number of texture elements pass by a reference point in a fixed visual direction per unit of time regardless of eyeheight. In Experiments 1 and 2, however, the grid that composed the ground surface was rescaled with eyeheight so that texture density was invariant. The consequence of scaling grid size with eyeheight was that GOFR and ER varied together. When eyeheight was halved, for example, both GOFR and ER were doubled. The results of Experiments 1 and 2 clearly indicate that braking is influenced by information about speed of self-motion, but they do not allow for the drawing of any conclusions about the separate contributions of GOFR and ER. In Experiment 3, eyeheight and ground-surface grid size were manipulated independently to determine how both optical variables contribute to braking.

Method
Participants. Twelve undergraduate students (6 women and 6 men) participated in the experiment. Their average age was 19.2 years, and all had normal or corrected-to-normal vision. All but 1 participant had a valid drivers license, and their mean number of years of driving experience was 3.2 (SD 2.0 years). Displays and apparatus. The ground texture used in Experiment 3 differed from the ground texture used in Experiments 1 and 2. Because the texture used in the previous experiments contained both coarse and finegrained detail (see Figure 3), discontinuities that were not salient at a small scale would become noticeable at a larger scale. If ER is determined by the discontinuities that are salient at each scale, then rescaling a texture with coarse and fine-grained detail may not necessarily change ER. So a brick texture that did not contain salient detail at nested scales was used in Experiment 3. Sample frames from the smallest and largest grid-size conditions are shown in Figure 10. There were three initial TTCs (2.5, 3.5, and 4.5 s), three eyeheights (0.55, 1.10, and 2.20 m), and three grid sizes (0.5 , 1.0 , and 2.0 ).7 On each trial, initial distance varied randomly between 40 and 50 m, and sign radius varied randomly between 0.25 and 0.50 m. Procedure. Prior to the experimental session, observers completed one block of 48 practice trials. Initial TTC, initial distance, and sign radius varied as they did in the experimental session, but eyeheight and grid size were fixed at the middle values. Design. The design for Experiment 3 was 3 (eyeheight) 3 (grid size) 3 (initial TTC). All variables were within subject, and there were 10 repetitions per condition for a total of 270 trials (in addition to the 48 practice trials).

Figure 11D) as well as the interactions were nonsignificant (all Fs 1). Initiation of braking. Mean dideal at the onset of braking is plotted as a function of initial TTC for each level of eyeheight in Figure 11E and for each level of grid size in Figure 11F. The main effects of eyeheight, F(2, 22) 17.23, p .01 (partial 2 .61), grid size, F(2, 22) 10.63, p .01 (partial 2 .49), and initial TTC, F(2, 22) 1,666.67, p .01, were all significant. None of the interactions were significant. Magnitude of brake adjustments. For initial brake adjustments, the main effects of eyeheight, F(2, 22) 17.65, p .01 (partial 2 .62), and grid size, F(2, 22) 5.62, p .05 (partial 2 .39), were both significant (see Figure 12A), but the main effect of initial TTC, F(2, 22) 1.41, p .27, was not. None of the interactions were significant. As in Experiment 2, subsequent brake adjustments were separated into adjustments that resulted in an increase in brake pressure and adjustments that resulted in a decrease. Again, there was a tendency to undershoot when decreasing brake pressure, and there was no indication that decreases were influenced by speed information. For brake adjustments resulting in an increase, however, the main effect of eyeheight, F(2, 22) 14.44, p .01 (partial 2 .57), was significant, but the main effect of grid scale, F(2, 22) 1.43, p .26 (partial 2 .12), and the Eyeheight Grid-Scale interaction (F 1) were not (see Figure 12B). In sum, the results of Experiment 3 indicate that both GOFR and ER influence the initiation and magnitude of brake adjustments, affecting final stopping distance. However, effect sizes indicate that the contribution of GOFR is considerably greater. This is inconsistent with Larish and Flach (1990), who found that ER had a stronger effect on perceptual judgments of speed. It is possible that the influence of ER was magnified in their study because the screen occupied such a small portion of the observers field of view8 and because the border of the screen was clearly visible in one condition. Such viewing conditions may enhance the effect of ER because the texture discontinuities disappear behind the border of the screen. ER may be less salient under the more natural viewing conditions used in the present study because the border of the screen is less noticeable and is in the periphery of the observers visual field. This explanation is consistent with the conclusion of Dyre (1997), who found that perceptual judgments of speed were affected by GOFR but not by ER when simulated self-motion was presented across a wide field of view.

Experiment 4
In Experiments 1 and 2, the stop signs were floating in the air rather than being anchored to the ground by a post. Floating stop signs were used because of a concern that the presence of posts would provide a salient cue to changes in eyeheight. However,
7 A grid size of 0.5 means that the length and the width of each tile composing the ground surface were one half of the length and the width of the tiles in the 1.0 condition. 8 Larish and Flach (1990) used two viewing conditions. In the uncontrolled viewing condition, the display occupied 47 of the observers field of view along the horizontal dimension. In the controlled viewing condition, observers viewed displays through a narrow tube that allowed them to see only a 7.2 circular area. In the present study, the displays occupied approximately 84 (height) 62 (width) of the observers field of view.

Results and Discussion

The data were entered into separate 3 (eyeheight) 3 (texture size) 3 (initial TTC) repeated measures ANOVAs. Constant error in final stopping distance. The main effects of eyeheight, F(2, 22) 20.13, p .01 (partial 2 .65), initial TTC, F(2, 22) 29.22, p .01 (partial 2 .73), and grid size, F(2, 22) 4.89, p .05 (partial 2 .31), were all significant. In addition, there were significant Initial TTC Eyeheight, F(4, 44) 7.97, p .01 (see Figure 11A), and Initial TTC Grid Size, F(4, 44) 3.43, p .05 (see Figure 11B), interactions. Variable error in final stopping distance. Only the main effect of initial TTC, F(2, 22) 11.66, p .01, was significant. The main effects of eyeheight (see Figure 11C) and grid size (see

CALIBRATION AND BRAKING

493

Figure 10. Sample frames from the display used in the 0.5 Experiment 3.

(A) and 2.0

(B) grid-size conditions of

objects are typically anchored to the ground plane in natural environments, and the height of an objects base in the visual field provides a robust cue to distance. Meng and Sedgwick (2001, 2002) recently demonstrated that observers misperceive distance when objects are not connected to the ground plane. They showed that under static viewing conditions, perceived distance is based on the optical contact of the base of the object with the ground

surface. So objects that were floating in the air were perceived to be farther away than they actually were. Although the displays used in Experiments 13 contained simulated self-motion toward the sign, one could argue on the basis of previous research that observers misperceived the depth of the floating signs and that the conclusions cannot be generalized to more natural conditions. In Experiment 4, braking behavior with and without signposts

494

FAJEN

Figure 11. Constant error in stopping distance (A and B), variable error in stopping distance (C and D), and mean ideal deceleration at the onset of braking (E and F) for Initial Time to Contact (TTC) Eyeheight and Initial TTC Grid-Size interactions in Experiment 3.

was compared to determine whether performance is affected under conditions in which stationary observers misperceive distance. Posts connecting the bottoms of the stop signs with the ground plane were absent in one block of trials and were present in the

other block. Because the point of optical contact is also influenced by eyeheight, and eyeheight was varied in some conditions used in Experiments 13, the effect of post was tested in both the fixed-EH and variable-EH conditions.

CALIBRATION AND BRAKING

495

variable EH was a between-subjects variable, and post, initial distance, and initial TTC were within-subject variables. Trials were blocked by post, and block order was counterbalanced across participants. Trials were presented in a completely random order within each block, and there were four repetitions per condition for a total of 200 trials (plus 100 practice trials).

Results and Discussion

Data were collapsed across initial distance and entered into separate 2 (post) 5 (initial TTC) ANOVAs for the analysis of constant and variable error in final stopping distance. Separate analyses were run for the fixed-EH and variable-EH conditions. Constant error in final stopping distance. In the fixed-EH condition, neither the main effect of post, F(1, 8) 1.54, p .25, nor the Post Initial TTC interaction (F 1) was significant (see Figure 13A). Likewise, in the variable-EH condition, both the main effect of post and the Post Initial TTC interaction were nonsignificant (Fs 1; see Figure 13B). Although nonsignificant results cannot be used to accept the null hypothesis (i.e., that there are no differences between the post and no-post conditions), it is possible to show how small the differences are by calculating the 95% confidence intervals around the difference between the mean final stopping distance in the no-post and post conditions. Figure 14 shows that the upper and lower bounds of the 95% confidence interval are close to 0 in both the fixed-EH and variable-EH conditions. This means that it is possible to conclude with reasonable certainty that differences between the no-post and post conditions, if they even exist, are quite small (less than 1 m in the fixed-EH condition and less than 2 m in the variable-EH condition). For the purposes of comparison, consider the extent to which distance would be overestimated under the current conditions if observers relied on the optical contact of the bottom of the sign with the ground plane. The proportion of actual distance by which perceived distance would be overestimated is given by the ratio of eyeheight to sign radius. For example, if eyeheight is 1.10 m and sign radius is 0.50 m (the largest sign radius used in the present experiments), then distance would be overestimated by more than twice the actual distance. Even at close distances (e.g., 5 m), this corresponds to a substantial degree of overestimation. Thus, if braking control was biased by distance estimates based on the optical contact of the sign with the ground plane, the difference between the no-post and post conditions would be considerably greater. Variable error in final stopping distance. Again, the main effect of post in both conditions (Fs 1) and the Post Initial TTC interaction in the fixed-EH condition, F(4, 32) 1.39, p .26, and in the variable-EH condition (F 1; see Figures 13C and 13D) were nonsignificant. In addition, the upper and lower bounds of the 95% confidence intervals were quite close to 0 in both conditions (see Figures 14C and 14D). Although observers in Meng and Sedgwick (2001, 2002) overestimated the depth of floating objects under static viewing conditions, the results of Experiment 4 confirm that performance on the braking task was essentially unaffected when posts were absent. The lack of a post effect is most interesting in the variable-EH condition because the point of optical contact depends on eyeheight. So even under conditions in which the amount of overestimation of depth for a stationary observer would vary with eyeheight, performance on the braking task was still essentially equivalent in the post and no-post conditions.

Figure 12. Scaled magnitude of initial brake adjustment (A) and subsequent brake adjustments (B) for each grid size in the 0.55-, 1.10-, and 2.20-m eyeheight conditions of Experiment 3.

Method
Participants. Eighteen undergraduate students (5 women and 13 men) participated in the experiment. Their average age was 18.9 years, and all had normal or corrected-to-normal vision. All participants had a valid drivers license, and their mean number of years of driving experience was 2.6 (SD 1.2 years). Half of the students were assigned to the fixed-EH condition, and the other half were assigned to the variable-EH condition. Displays and apparatus. In the post condition, the stop signs were anchored to the ground plane by a white post. In the no-post condition, the stop signs were floating in air, as in Experiments 13. There were five initial distances (40.0, 42.5, 45.0, 47.5, and 50.0 m) and five initial TTCs (2.5, 3.0, 3.5, 4.0, and 4.5 s). Sign radius varied randomly on each trial between 0.25 and 0.50 m. Eyeheight was fixed at 1.10 m in the fixed-EH condition and varied randomly between 0.55 and 2.20 m in the variable-EH condition. As in Experiments 1 and 2, the size of the ground-plane grid was scaled with eyeheight. Also, the same texture that was used in Experiments 1 and 2 was used here. Procedure. The procedures were similar to those used in Experiment 1, with a few exceptions. Prior to the experimental session, observers completed one block of 100 practice trials. Observers who were assigned the post condition during the first experimental block also received the post condition during the practice session, and those who were assigned the no-post condition in the first experimental block also received the no-post condition during the practice session. Design. The design for Experiment 4 was 2 (fixed EH vs. variable EH) 2 (post) 5 (initial distance) 5 (initial TTC). Fixed EH versus

496

FAJEN

Figure 13. Constant (A and B) and variable (C and D) error in stopping distance as a function of initial time to contact for the post and no-post conditions of Experiment 4. EH eyeheight.

There are at least two possible explanations for why performance is unaffected when the signs are floating rather than anchored to the ground. First, although distance judgments appear to be based on the point of optical contact under static viewing conditions, it is unlikely that observers in the present study misperceived the floating stop signs as resting on the ground, because the point of optical contact changes as an observer approaches a floating sign. If the sign were resting on the ground, the point of optical contact would be the same. Second, the visual control of braking is based on optical variables that depend on distance, but observers do not need to perceive distance to control braking. Lastly, recall that the decision to use floating stop signs in Experiments 13 was taken due to a concern that posts would provide a salient cue to eyeheight that observers would be able to use to compensate for the effect of changing eyeheight on GOFR. The results of Experiment 4 suggest that this concern was needless. Comparison of the fixed-EH and variable-EH conditions in Figures 13 and 14 indicates that the effect of eyeheight was robust even when posts were present. It appears that observers were

unable to compensate for the effect of changing eyeheight on GOFR by using the length of the posts as a cue to eyeheight.

General Discussion
One goal of this study was to test existing models of braking to avoid a collision by exploring the role of information about speed of self-motion provided by GOFR and ER. Experiment 1 extended Yilmaz and Warrens (1995) study on actively controlled braking by comparing performance under rapid braking conditions when GOFR and ER were present, absent, and unreliable. GOFR and ER were biased by randomly varying the simulated eyeheight of the observer above a textured ground surface and by rescaling the size of the grid composing the ground surface with eyeheight. Together, the manipulations of eyeheight and grid size altered the relation between speed and GOFR and that between speed and ER. The results indicated that mean error in final stopping distance, variable error in final stopping distance, and percentage of crashes were all affected when GOFR and ER were absent or unreliable,

CALIBRATION AND BRAKING

497

Figure 14. Difference in mean final stopping distance (A and B) and difference in standard deviation of final stopping distance (C and D) between the no-post and post conditions as a function of initial time to contact in Experiment 4. Vertical bars indicate 95% confidence intervals.

especially when initial TTC was short. Observers tended to stop short of the target when GOFR and ER were higher than normal, and they were more likely to crash into the target when GOFR and ER were lower than normal. In Experiment 2, eyeheight and grid size were again manipulated together, but this time as a withinsubject independent variable to allow for more detailed analyses of individual braking adjustments. These analyses demonstrated that GOFR and ER influenced the initiation of braking and the magnitude of brake adjustments. When GOFR and ER were increased by halving eyeheight and grid size, observers initiated braking earlier and made larger than normal brake adjustments. When eyeheight and grid size were manipulated independently in Experiment 3, it was found that both GOFR and ER influence braking but that GOFR has a stronger effect. Lees (1976) model cannot account for the influence of GOFR and ER because the optical information in this model is defined by the optical angle of the target (and its derivatives) and is independent of optic flow from the ground plane. The deceleration error model can explain the effects of GOFR and ER because these variables are used to estimate dideal (see Equation 2). However, neither model can account for observers sensitivity to the action boundary defined by Dmax. In Experiment 1, such sensitivity was demonstrated by showing that observers reliably increase brake pressure when dideal approaches Dmax, even when the deceleration

error is close to 0 (i.e., when is close to 0.5). To account for these findings, the error-nulling assumption on which the and deceleration error models are based was rejected, and an entirely new kind of modelthe affordance modelwas introduced. According to the affordance model, observers make brake adjustments to keep dideal within a safe region greater than 0 and less than Dmax. This model accounts for the fact that observers do not always initiate braking immediately on recognition of an obstacle or make small, frequent brake adjustments throughout the approach to cancel the difference between dcurrent and dideal. Deceleration errors can be safely tolerated for brief periods, as long as dideal does not exceed Dmax. When dideal does approach Dmax, however, observers must increase brake pressure immediately even if the deceleration error is small. This explains why observers in Experiment 1 reliably increased brake pressure when dideal was near Dmax. In addition, the new model also allows for individual differences in the style of controlwhereas conservative drivers keep dideal well below Dmax, aggressive drivers often allow dideal to get as close as possible to Dmax before braking. To make the distinction between the error-nulling models and the new affordance model more concrete, consider the property of the actor environment system that must be perceived to control braking. According to the and deceleration error models, the relevant property is I am (not) braking hard enough to stop. The

498

FAJEN

question then becomes one of how observers are informed of this property. The optical variable is one possible answer, and the difference between dcurrent and dideal is another. According to the affordance model, the relevant property is it is (not) still possible to stop. When the observer is far away from the obstacle or is moving slowly, dideal is less than Dmax, and it is still possible to stop. But if the observer approaches the obstacle without braking, then there is some point at which dideal exceeds Dmax, and it is no longer possible to stop. I am (not) braking hard enough to stop and it is (not) still possible to stop are different properties that are specified by different optical variables. Whether it is still possible to stop depends on dideal relative to Dmax. This means that researchers should be looking for optical variables that specify dideal rather than optical variables that specify the sufficiency of dcurrent. Equation 2 shows that dideal is optically specified by a combination of GOFR or ER and (i.e., dideal GOFR/2 , or dideal ER/2 ). The effects of GOFR and ER observed in Experiments 13 are consistent with use of these variables. When eyeheight was doubled, for example, observers waited longer before increasing brake pressure, and they made smaller brake adjustments because perceived dideal based on GOFR/2 was underestimated. However, it is important to note that there are a variety of other combinations of optical variables that correlate with, but do not uniquely specify, dideal. The present data allow for a conclusion that perceived dideal was based in part on GOFR (and, to a lesser extent, ER), but they do not rule out the possibility that observers used a nonspecifying variable rather than GOFR/2 . Previous research suggests that different observers may rely on different optical variables, especially during the early stages of learning. With practice and experience, observers may converge on higher order specifying variables, such as GOFR/2 . The transition from using one optical variable to another is also likely to depend on the range of conditions experienced by the observer (e.g., slow or fast conditions), the presence of local constraints (e.g., fixed sign size), and the criteria for success (Jacobs, Runeson, & Michaels, 2001; Michaels & de Vries, 1998; Pagano & Bingham, 1998; Runeson, Juslin, & Olsson, 2000; Smith, Flach, Dittman, & Stanard, 2001; Tresilian, 1999). Fajen and Devaney (2005) recently investigated the use of nonspecifying variables to control braking by asking observers to wait until the last possible moment before initiating maximum brake pressure. To perform the task, observers should initiate braking when perceived dideal Dmax. If observers rely on a specifying variable such as GOFR/2 , dideal at the onset of braking should be unaffected by factors such as sign size and initial speed. However, biases due to sign size and initial speed were present, suggesting that observers do use nonspecifying variables.

The Necessity of Calibration

Experiments 13 demonstrated that GOFR and (to a lesser extent) ER are used to control braking, which is what one would expect if observers detect information about dideal. However, whether it is still possible to stop also depends on Dmax or the strength of the brake. Clearly, there is no optical information about Dmax. Furthermore, the effective strength of the brake can change due to a variety of factors, including changes in controller dynamics, load, surface friction, or surface slope. So if successful braking is performed by keeping dideal within the safe region, and the boundary of that region defined by the maximum strength of the

brake is invisible, how do observers know the location of that boundary? One might assume that observers must perceive dideal and then compare it with knowledge retrieved from memory about the maximum deceleration of the brake. The problem with this account is that when two quantities are compared, they must be in the same units. Decelerations are typically represented in units of D/T2 (i.e., distance per unit of time squared), so perhaps perceived dideal and knowledge of Dmax are both converted into eyeheight/s2. But seconds (and, hence, eyeheight/s2) are extrinsic units that are convenient for the experimenter but arbitrary and meaningless to an observer. Rather than converting perceived dideal into extrinsic units, observers may learn to detect information about dideal in intrinsic units of Dmax through a process of calibration (Dmax is an intrinsic unit because it is defined by the observers action capabilities). This eliminates the need for storing and retrieving knowledge of Dmax in memory and the need for a process of conversion of dideal and Dmax into common units for comparison. For a well-calibrated observer, dideal 1.0 (in units of Dmax) means that it is still possible to stop within the limits of the braking capabilities. When one says that a novice does not know the strength of the brake, what it means in these terms is that the observer does not have the correct units. So he or she might encounter a situation in which dideal 1.0 but it is impossible to stop even when maximum brake pressure is applied. The novice must use information contained in the optical consequences of individual brake adjustments to adjust the size of the units until dideal 1.0 reliably separates situations in which it is possible to stop from situations in which it is no longer possible to stop.9 And the same applies to situations in which the effective strength of the brake changes due to factors such as changing load (e.g., towing a trailer), surface friction (e.g., driving on ice), or surface slope. Just as humans are capable of recalibrating locomotion to changes in the amount of effort required to walk (Rieser, Pick, Ashmead, & Garing, 1995), one would expect them to recalibrate to changes in the maximum deceleration of the brake. The issue of calibration was recently investigated by Fajen (in press-b) by manipulating the maximum deceleration of the brake. Recent research by Proffitt and colleagues (Bhalla & Proffitt, 1999; Proffitt, Bhalla, Gossweiler, & Midgett, 1995; Proffitt, Stefanucci, Banton, & Epstein, 2003) has provided evidence that spatial properties are perceived in intrinsic units defined by observers capabilities to act. This research has demonstrated that judgments of distance and surface slope are influenced by factors that affect the capabilities of the observer to walk or climb (e.g., wearing a backpack, being tired or in poor physical condition). Accordingly, one could argue that the process of calibration results in the detection of information about dideal in units of the observers capabilities for action, which in the case of braking are defined by Dmax. It is interesting to note that during the debriefing sessions for the experiments in which eyeheight was manipulated, several observers reported that the strength of the brake seemed to change from trial to trial. If the strength of the brake is measured in terms
In practice, observers are likely to calibrate to the maximum rate of deceleration that feels safe and comfortable rather than the maximum deceleration of the brake. The conservative approach that novices are likely to adopt is to begin with very small units, so dideal will never come close to Dmax.
9

CALIBRATION AND BRAKING

499

of the values of perceived dideal (based on GOFR/2 ) that can be controlled, then the brake was effectively twice as strong in the low-eyeheight condition and half as strong in the high-eyeheight condition. That is, in units that are relevant to the observer, changing eyeheight effectively changes the strength of the brake.

Styles of Control and the Affordance Model

Yilmaz and Warren (1995) classified trials into categories (e.g., regulated, bang-bang, slam-on-the-brakes) based on the style of control. Although the large majority of trials (83%) were classified as regulated, 11% were classified as bang-bang, and 6% were classified as slam-on-the-brakes, suggesting that deceleration profiles are not always characterized by small, frequent brake adjustments. Yilmaz and Warren were forced to classify trials because the analyses that they performed could only be run on regulated trials. So they ended up excluding 17% of the trials, regardless of whether a trial resulted in a successful stop. Because analyses were only performed on regulated trials, one is led to conclude that regulated trials resulted from a control strategy, whereas bangbang and slam-on-the-brakes trials resulted from some other strategy. According to the affordance model, however, these styles are not the result of different control strategies but, rather, the result of differences among observers in the significance placed on certain costs (e.g., approach time, collision rate). For example, observers could minimize the risk of collision by applying hard braking early in the approach (i.e., while dideal is still low). This would correspond to what Yilmaz and Warren called bang-bang control. Alternatively, observers could minimize approach time by waiting until the last moment (i.e., until dideal is almost equal to Dmax) to apply maximum brake pressure, which would correspond to a slam-on-the-brakes trial. Another option would be to make small, frequent brake adjustments to keep dideal within the middle of the safe region, which would be classified as a regulated trial. Thus, although there are many ways in which one could perform the braking task successfully, these differences simply result from differences in the significance placed on certain costs. There is no need to exclude trials because the control strategy is the same regardless of the importance of minimizing approach time or collision rate, observers always make brake adjustments to keep dideal within the safe region. Different styles simply correspond to different trajectories through the safe region.

are body scaled because they are defined by dimensions of the environment taken with respect to dimensions of the body, such as leg and arm length or shoulder width. In contrast, the affordances in visually guided actions are defined by ones action capabilities (Oudejans, Michaels, Bakker, & Dolne, 1996). Whether it is still possible to stop depends on Dmax, not on leg length or shoulder width. Thus, if observers detect information about dideal in units of Dmax, braking is controlled by perceiving whether it is possible to stopthat is, braking is controlled by perceiving an affordance. Finally, a variety of other perceptionaction tasks have been modeled as processes by which humans null the error between the current and the ideal values of a perceptual variable (see Fajen, in press-a). Most steering models are based on the idea that the observer nulls the heading error between the direction of motion and the goal (Fajen, 2001; Wann & Land, 2000; Warren, Kay, Zosh, Duchon, & Sahuc, 2001; Wilkie & Wann, 2003). The required velocity model (Peper, Bootsma, Mestre, & Bakker, 1994) of intercepting a moving target by hand is based on the idea that the observer nulls the lateral position error between the hand and the ball before the ball crosses the plane of the body. The optical acceleration cancellation (Chapman, 1968; Michaels & Oudejans, 1992) and linear optical trajectory (McBeath, Shaffer, & Kaiser, 1995) models of fly ball catching assume that the observer nulls the optical acceleration or optical curvature of the ball. As in braking, the urgency of steering wheel adjustments or changes in hand speed and running speed depend on the maximum capabilities of the observer. Thus, sensitivity to the limits of ones action capabilities is just as likely to play an essential role in these other perceptionaction tasks, and existing error-nulling models of steering, intercepting moving objects, and catching fly balls should be reformulated as affordance models.

References
Andersen, G. J., Cisneros, J., Atchley, P., & Saidpour, A. (1999). Speed, size, and edge-rate information for the detection of collision events. Journal of Experimental Psychology: Human Perception and Performance, 25, 256 269. Andersen, G. J., & Sauer, C. W. (2004). Optical information for collision detection during deceleration. In H. Hecht & G. J. P. Savelsbergh (Eds.) Time to contact: Advances in psychology (Vol. 135, pp. 93108). Amsterdam: Elsevier. Ballard, T. G., Roach, T., & Dyre, B. P. (1998). Use of global optical flow rate and discontinuity rate depends on their validity as determinants of egospeed. In Proceedings of the 42nd Annual Meeting of the Human Factors and Ergonomics Society (pp. 1440 1444). Santa Monica, CA: Human Factors and Ergonomics Society. Bardy, B. G., & Laurent, M. (1998). How is body orientation controlled during somersaulting? Journal of Experimental Psychology: Human Perception and Performance, 24, 963977. Bardy, B. G., & Warren, W. H. (1997). Visual control of braking in goal-directed action and sport. Journal of Sport Sciences, 15, 607 620. Bhalla, M., & Proffitt, D. R. (1999). Visualmotor recalibration in geographical slant perception. Journal of Experimental Psychology: Human Perception and Performance, 25, 1076 1096. Bootsma, R. J., & van Wieringen, P. C. W. (1990). Timing an attacking forehand drive in table tennis. Journal of Experimental Psychology: Human Perception and Performance, 16, 2129. Calderone, J. B., & Kaiser, M. K. (1989). Visual acceleration detection: Effect of sign and motion orientation. Perception & Psychophysics, 45, 391394. Carello, C., Grosofsky, A., Reichel, F. D., Solomon, H. Y., & Turvey,

Affordances and Visually Guided Braking

The new model of visually guided braking introduced in this article is called the affordance model because it suggests that braking is controlled by making adjustments so that safe stopping is always afforded within the limits of ones braking capabilities. This means that observers must be able to reliably perceive whether it is still possible to stop. The view that successful performance depends on the ability to perceive which actions are possible and which actions are not possible has its origins in Gibsons (1979) theory of affordances. From the Gibsonian perspective, humans and animals control their interactions with the environment by perceiving affordances (Turvey, 1992). Most affordance research has focused on body-scaled affordances such as stair climbing (Mark, 1987; Warren, 1984), reaching (Carello, Grosofsky, Reichel, Solomon, & Turvey, 1989), and passing through an aperture (Warren & Whang, 1987). These affordances

500

FAJEN through nested contact relations among surfaces. Perception & Psychophysics, 63, 115. Meng, J. C., & Sedgwick, H. A. (2002). Distance perception across spatial discontinuities. Perception & Psychophysics, 64, 114. Michaels, C. F., & de Vries, M. M. (1998). Higher order and lower order variables in the visual perception of relative pulling force. Journal of Experimental Psychology: Human Perception and Performance, 24, 526 546. Michaels, C. F., & Oudejans, R. R. D. (1992). The optics and actions of catching fly balls: Zeroing out optical acceleration. Ecological Psychology, 4, 199 222. Michaels, C. F., Zeinstra, E. B., & Oudejans, R. R. D. (2001). Information and action in punching a falling ball. Quarterly Journal of Experimental Psychology: Human Experimental Psychology, 54(A), 59 93. Moen, G. C., DiCarlo, D. J., & Yenni, K. R. (1976). A parametric analysis of visual approaches for helicopters (NASA Tech. Note D-8275). Langley, VA: National Aeronautics and Space Administration. Oudejans, R. R. D., Michaels, C. F., Bakker, F. C., & Dolne, M. A. (1996). The relevance of action in perceiving affordances: Perception of catchableness of fly balls. Journal of Experimental Psychology: Human Perception and Performance, 22, 879 891. Pagano, C. C., & Bingham, G. P. (1998). Comparing measures of monocular distance perceptions: Verbal and reaching errors are not correlated. Journal of Experimental Psychology: Human Perception and Performance, 24, 10371051. Peper, C. E., Bootsma, R. J., Mestre, D. R., & Bakker, F. C. (1994). Catching balls: How to get the hand in the right place at the right time. Journal of Experimental Psychology: Human Perception and Performance, 20, 591 612. Proffitt, D. R., Bhalla, M., Gossweiler, R., & Midgett, J. (1995). Perceiving geographical slant. Psychonomic Bulletin & Review, 2, 409 428. Proffitt, D. R., Stefanucci, J., Banton, T., & Epstein, W. (2003). The role of effort in perceiving distance. Psychological Science, 14, 106 112. Rieser, J. J., Pick, H. L., Ashmead, D. H., & Garing, A. E. (1995). Calibration of human locomotion and models of perceptualmotor organization. Journal of Experimental Psychology: Human Perception and Performance, 21, 480 497. Runeson, S., Juslin, P., & Olsson, H. (2000). Visual perception of dynamic properties: Cue heuristics versus direct-perceptual competence. Psychological Review, 107, 525555. Rushton, S. K., & Wann, J. P. (1999). Weighted combination of size and disparity: A computational model for timing a ball catch. Nature Neuroscience, 2, 186 190. Saidpour, A. & Andersen, G. J. (in press). The role of constant deceleration in the control of vehicle braking. In A. G. Gale (Ed.), Vision in vehicles IX. Amsterdam: Elsevier Science. Savelsbergh, G. J. P., Whiting, H. T. A., & Bootsma, R. J. (1991). Grasping tau. Journal of Experimental Psychology: Human Perception and Performance, 17, 315322. Smith, M. R. H., Flach, J. M., Dittman, S. M., & Stanard, T. (2001). Monocular optical constraints on collision control. Journal of Experimental Psychology: Human Perception and Performance, 27, 395 410. Spurr, R. T. (1969). Subjective aspects of braking. Automobile Engineering, 59, 58 61. Tresilian, J. R. (1999). Visual timed action: Time-out for tau? Trends in Cognitive Sciences, 3, 301309. Turvey, M. T. (1992). Affordances and prospective control: An outline of the ontology. Ecological Psychology, 4, 173187. Wann, J. P., Edgar, P., & Blair, D. (1993). Time-to-contact judgment in the locomotion of adults and preschool children. Journal of Experimental Psychology: Human Perception and Performance, 19, 10531065. Wann, J. P., & Land, M. F. (2000). Steering with or without the flow: Is the retrieval of heading necessary? Trends in Cognitive Sciences, 4, 319 324. Warren, W. H. (1984). Perceiving affordances: Visual guidance of stair

M. T. (1989). Visually perceiving what is reachable. Ecological Psychology, 1, 2754. Chapman, S. (1968). Catching a baseball. American Journal of Physics, 53, 849 855. Craig, C. M., & Bootsma, R. J. (2001). The role of tau-dot information in detecting upcoming collision. In G. A. Burton & R. C. Schmidt (Eds.), Studies in perception and action VI (pp. 77 80). Mahwah, NJ: Erlbaum. DeLucia, P. R., & Warren, R. (1994). Pictorial and motion based information during active control of self-motion: Size-arrival effects on collision avoidance. Journal of Experimental Psychology: Human Perception and Performance, 20, 783789. Dyre, B. P. (1997). Perception of accelerating self-motion: Global optical flow rate dominates discontinuity rate. In Proceedings of the 41st Annual Meeting of the Human Factors and Ergonomics Society (pp. 1333 1337). Santa Monica, CA: Human Factors and Ergonomics Society. Fajen, B. R. (2001). Steering toward a goal by equalizing taus. Journal of Experimental Psychology: Human Perception and Performance, 27, 953968. Fajen, B. R. (in press-a). Perceiving possibilities for action: On the necessity of calibration and perceptual learning for the visual guidance of action. Perception. Fajen, B. R. (in press-b). Scaling information to action in visually guided braking. Journal of Experimental Psychology: Human Perception and Performance. Fajen, B. R., & Devaney, M. C. (2005). Learning to control collisions: The role of perceptual attunement and action boundaries. Manuscript submitted for publication. Gibson, J. J. (1979). The ecological approach to visual perception. Boston: Houghton Mifflin. Gray, R., & Regan, D. (1998). Accuracy of estimating time to collision using binocular and monocular information. Vision Research, 38, 499 512. Harris, L. R., Jenkin, M., Zikovitz, D., Redlick, F., Jaekl, P., Jasiobedzka, U., et al. (2002). Simulating self motion: I. Cues for the perception of motion. Virtual Reality, 6, 75 85. Jacobs, D. M., Runeson, S., & Michaels, C. F. (2001). Learning to visually perceive the relative mass of colliding balls in globally and locally constrained task ecologies. Journal of Experimental Psychology: Human Perception and Performance, 27, 1019 1038. Kaiser, M. K., & Phatak, A. V. (1993). Things that go bump in the light: On the optical specification of contact severity. Journal of Experimental Psychology: Human Perception and Performance, 19, 194 202. Kim, N.-G., Turvey, M. T., & Carello, V. (1993). Optical information about the severity of upcoming contacts. Journal of Experimental Psychology: Human Perception and Performance, 19, 179 193. Larish, J. F., & Flach, J. M. (1990). Sources of optical information useful for perception of speed of rectilinear self-motion. Journal of Experimental Psychology: Human Perception and Performance, 16, 295302. Lee, D. N. (1976). A theory of visual control of braking based on information about time-to-collision. Perception, 5, 437 459. Lee, D. N., Davies, M. N. O., Green, P. R., & van der Weel, F. R. (1993). Visual control of velocity of approach by pigeons when landing. Journal of Experimental Biology, 180, 85104. Lee, D. N., Reddish, P. E., & Rand, D. (1991). Aerial docking by hummingbirds. Naturwissenschaften, 78, 526 527. Lee, D. N., Young, D. S., & Rewt, D. (1992). How do somersaulters land on their feet? Journal of Experimental Psychology: Human Perception and Performance, 18, 11951202. Mark, L. S. (1987). Eyeheight-scaled information about affordances: A study of sitting and stair climbing. Journal of Experimental Psychology: Human Perception and Performance, 13, 361370. McBeath, M. K., Shaffer, D. M., & Kaiser, M. K. (1995, April 28). How baseball outfielders determine where to run to catch fly balls. Science, 268, 569 573. Meng, J. C., & Sedgwick, H. A. (2001). Distance perception mediated

CALIBRATION AND BRAKING climbing. Journal of Experimental Psychology: Human Perception and Performance, 10, 683703. Warren, W. H., Kay, B. A., Zosh, W. D., Duchon, A. P., & Sahuc, S. (2001). Optic flow is used to control human walking. Nature Neuroscience, 4, 213216. Warren, W. H., & Whang, S. (1987). Visual guidance of walking through apertures: Body-scaled information for affordances: Journal of Experimental Psychology: Human Perception and Performance, 13, 371383. Wilkie, R., & Wann, J. (2003). Controlling steering and judging heading: Retinal flow, visual direction, and extraretinal information. Journal of

501

Experimental Psychology: Human Perception and Performance, 29, 363378. Yilmaz, E. H., & Warren, W. H. (1995). Visual control of braking: A test of the hypothesis. Journal of Experimental Psychology: Human Perception and Performance, 21, 996 1014.

Received May 8, 2003 Revision received June 29, 2004 Accepted November 8, 2004